Flux density distribution forecasting in concentrated solar tower plants: A data-driven approach

Concentrated Solar Power (CSP) systems, particularly those employing heliostat fields combined with a central tower, demonstrate substantial capacity for producing dispatchable, sustainable energy and fuel. This is achieved by focusing the sunlight with up to thousands of individual heliostats onto a single receiver. Forecasting the focal spot of each heliostat at any solar position becomes imperative to ensure optimal control. Nevertheless, the existing cutting-edge techniques aimed at predicting this flux density distribution either suffer from inaccuracies or entail substantial costs. In response to these challenges, our study introduces a novel approach involving a generative model that learns the shape and intensity patterns of the focal spots directly from images captured of the calibration target. We developed a purely data-driven methodology to generate the focal spots of the heliostats corresponding to various sun positions. The model is based on the StyleGAN architecture with adapted learnable input vectors for each individual heliostat and sun positions as input condition. The methodology’s effectiveness is demonstrated through training and evaluation on data collected from a research power plant, where it achieved a flux prediction accuracy of 89% on the calibration target surface. Our work offers a novel solution for predicting flux density distributions in solar power plants in a fully data-driven way with a neural network. This method achieves cost efficiency by utilizing data obtained during standard operational procedures. Impressively, this method attains accuracy levels comparable to or exceeding those of current state-of-the-art techniques. • Developed a novel, data-driven approach leveraging a modified StyleGAN architecture. • Achieved 90% accuracy in predicting heliostats focal spots. • Used purely data-driven techniques with camera-target images, eliminating physical models. • Demonstrated significant improvements over traditional flux prediction methods. • Showcased the model’s efficiency and scalability for large heliostat fields with limited data.